Private beta · Early access
Reference
API reference
Request fields, limits, and error codes for `POST /v1/chat/completions`.
Private beta. InferenceIndex is changing fast. Models, endpoints, pricing and these docs may change without notice, so check back before you build on a detail.
The call is the OpenAI-compatible chat completion from the quickstart. This page is the fields, limits, and error codes on that request. Saving a skill and naming one are on the skills page.
Request fields
reasoning_effort
How hard your fallback model reasons on the steps it writes. Omit it and the effort is none.
reasoning_effort is one of none, minimal, low, medium, high, xhigh, or max.
You can send the same choice as reasoning: { "effort": "low" }. reasoning.enabled set to true means medium. Set to false it means none. If reasoning_effort and reasoning.effort are both set and they differ, the request is refused with invalid_reasoning.
The response reports the effort it used as inferenceindex.reasoning_effort, or none when the request asked for none.
today
today is a date, YYYY-MM-DD. Steps that depend on the date use it. Omit it and the server's date is used.
stream_options.include_usage
stream: true streams the completion. Tool calls still arrive whole. Set stream_options to { "include_usage": true } to add a final chunk that carries usage and inferenceindex. Without it, inferenceindex is on the last content chunk and that stream has no usage chunk.
Model
The model is inferenceindex-0.1. A saved skill is inferenceindex-0.1/skl_ plus 32 hex characters, and @ plus a version from 1 pins that version: inferenceindex-0.1/skl_…@2. The same id can be sent as skill_id and skill_version in the body, or as X-Skill-Id and X-Skill-Version. The skills page has the rest. GET /v1/models lists inferenceindex-0.1 and each of your saved skills.
Limits
| Limit | What is refused |
|---|---|
| 500 tools | A request names at most 500 tools. |
| 100 tool calls | A run holds at most 100 tool calls. More is history_too_long. |
| 100,000 characters | The user messages, taken together, and the system and developer messages, taken together, are each at most 100,000 characters. More is too_long. |
A skill is at most 20,000 characters. That limit is on the skills page.
Errors
These chat completion errors use the OpenAI error shape. type is invalid_request_error. param is set when one field is at fault.
{
"error": {
"message": "Send the task as a user message",
"type": "invalid_request_error",
"code": "request_required"
}
}| Code | Status | When |
|---|---|---|
invalid_settings | 400 | settings was not accepted. param names the setting. |
invalid_skill_id | 400 | X-Skill-Id is not skl_ and 32 hex characters. |
invalid_skill_version | 400 | X-Skill-Version is not a whole number from 1. |
history_too_long | 400 | The run has more than 100 tool calls. The message is A run is limited to 100 tool calls. |
too_long | 400 | User messages, or system and developer messages, are over 100,000 characters. The message is User and system messages are limited to 100,000 characters each. |
tool_result_missing | 400 | A tool call in the history has no tool message. |
request_required | 400 | There is no user message. The message is Send the task as a user message. An image with no text still counts as a user message. |
invalid_reasoning | 400 | reasoning_effort and reasoning.effort ask for different efforts. |
insufficient_credit | 402 | The account has no credit left. Nothing runs. The message is The account has no credit left. |
401 and 403 are the same refusals as on the API keys page: a wrong or missing key, or an account that has not redeemed an invite or is disabled.
429 means the account already has 4 requests running. The body is { "error": "too_many_requests" }, the response has a Retry-After header, and the request is not charged. The same limit is on the credits page.